openai-cookbook icon indicating copy to clipboard operation
openai-cookbook copied to clipboard

Published 20 hours ago • verified publisher icon openai

Can customized loss be supported in fine-tuning

Open ccsquare opened this issue 2 years ago • 1 comments
trafficstars

Hi there,

It seems that only classification tasks and generation tasks can be fine-tuned based on fine-tunes API. Am I right? Is it possible for you to support much more fine-tuning tasks like reward model in paper Learning to summarize from human feedback which is implemented with pairwise loss.

ccsquare avatar Mar 10 '23 05:03 ccsquare

I have the same problem

huangguanMayday avatar Mar 10 '23 06:03 huangguanMayday

That's right. We do not support that today.

ted-at-openai avatar Mar 20 '23 18:03 ted-at-openai